Continual Learning via Local Module Composition

By Oleksiy Ostapenko, Pau Rodr'iguez L'opez, Massimo Caccia, Laurent Charlin. 2021

93 citations recorded by the source corpus.

Modularity is a compelling solution to continual learning (CL), the problem of modeling sequences of related tasks. Learning and then composing modules to solve different tasks provides an abstraction to address the principal challenges of CL including catastrophic forgetting, backward and forward transfer across tasks, and sub-linear model growth. We introduce local module composition (LMC), an approach to modular CL where each module is provided a local structural component that estimates a module's relevance to the input. Dynamic module composition is performed layer-wise based on local relevance scores. We demonstrate that agnosticity to task identities (IDs) arises from (local) structural learning that is module-specific as opposed to the task- and/or model-specific as in previous works, making LMC applicable to more CL settings compared to previous works. In addition, LMC also tracks statistics about the input distribution and adds new modules when outlier samples are detected. In the first set of experiments, LMC performs favorably compared to existing methods on the recent Continual Transfer-learning Benchmark without requiring task identities. In another study, we show that the locality of structural learning allows LMC to interpolate to related but unseen tasks (OOD), as well as to compose modular networks trained independently on different task sequences into a third modular network without any fine-tuning. Finally, in search for limitations of LMC we study it on more challenging sequences of 30 and 100 tasks, demonstrating that local module selection becomes much more challenging in presence of a large number of candidate modules. In this setting best performing LMC spawns much fewer modules compared to an oracle based baseline, however, it reaches a lower overall accuracy. The codebase is available under https://github.com/oleksost/LMC.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Meta & Continual Learning

Citations in the corpus

A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “MNTDP [245] and LMC [246] seek to explicitly ﬁnd the optimal layout from previous sub-modules and potentially new sub-modules.”
Learnability and Algorithm for Continual Learning (2023, background): “LMC (Ostapenko et al., 2021) learns task specific knowledge via local modules capable of task-id prediction.”
New Insights on Reducing Abrupt Representation Change in Online Continual Learning (2021, background): “The single-head assumption is further studied in task-agnostic continual learning settings (He et al., 2019; Caccia et al., 2020; Ostapenko et al., 2021; Von Oswald et al., 2021) in which the task-boundary assumption, amongst others, is also relaxed.”
Continual Learning with Foundation Models: An Empirical Study of Latent Replay (2022, methodology): “…Aljundi et al., 2017), dynamic architectures based methods (Yoon et al., 2017; Schwarz et al., 2018), which include modular methods (Mendez and EATON, 2021; Veniat et al., 2021; Ostapenko et al., 2021), and constrained optimization based methods (Lopez-Paz and Ranzato, 2017; Chaudhry et al., 2018).”
Continual Learning, Fast and Slow (2021, methodology): “Note that there is a hybrid setting with task identifiers provided during training but not during evaluation [3], which we will consider as task-aware in this work.”
Class-Incremental Learning: A Survey (2023, background): “On the other hand, several works [152], [186] propose to design specific modules for incremental new tasks, e.g. , adapters [187].”
How to Reuse and Compose Knowledge for a Lifetime of Tasks: A Survey on Continual Learning and Functional Composition (2022, background): “As examples, Mendez et al. (2020) demonstrated an improvement of 82.5% over noncompositional methods, and Ostapenko et al. (2021) demonstrated zero-shot compositional generalization to novel task combinations.”
Deep Class-Incremental Learning: A Survey (2023, background): “On the other hand, several works [153], [154], [155] propose to design specific modules for incremental new tasks, e.”
A Probabilistic Framework for Modular Continual Learning (2023, methodology): “Modular approaches to CL [41, 44, 30] represent a deep neural network as a composition of modules, each of which is a parameterized nonlinear transformation, such as a multilayer neural network.”
SketchOGD: Memory-Efficient Continual Learning (2023, unknown): “Some approaches are based on rehearsing old data/experiences whilst learning new tasks [12–17] or allocating neural resources incrementally for new tasks [18–21], while others rely on regularizing training to preserve performance on earlier tasks [22].”
A Model or 603 Exemplars: Towards Memory-Efficient Class-Incremental Learning (2022, background): “(Ostapenko et al., 2021) expands dynamic modules with dynamic routing to generalize to related tasks.”